Vehicle Counting Quantitative Comparison Using Background Subtraction, Viola Jones and Deep Learning Methods

In this paper, vehicle counting is investigated using various machine methods on four datasets. Vehicle counting is needed to complete the data required for short term predictions using highway traffic model, which is in turn, applicable for road design and usage planning. The goal of this research is to show that automatic car counting using machine methods, can be obtained from utilizing existing CCTV image data or from better cameras. Then by applying quantitative evaluation, F1 and precision scores are obtained, so that a few recommendations can be given. Through numerous simulations, F1 scores ranging from 0.32 to 0.75 have been successfully obtained for one low resolution dataset, using Background Subtraction and Viola Jones methods, on existing CCTV data. It has been found also that Viola Jones method can improve F1 score, by about 39% to 56%, over Back Subtraction method. Furthermore, the use of Deep Learning especially YOLO has provided good results, with F1 scores ranging from 0.94 to 1 and its precision ranges from 97.37% to 100% involving three datasets of higher resolution.

I. Introduction

Highway vehicle traffic management and planning are not common in developing countries. This is because they require integrated Intelligent Transport System (ITS) to collect vehicle data. A few approaches do exist [1], but in Indonesia traffic data are only made available from the existing Close Circuit Television (CCTV), from radio stations' road condition news and from police twitter. They are only raw, consequently, it would be great if these data can be made useful [2]. CCTV data are made available to the public freely by a company called Jasa Marga, operating in Indonesia [3], and are installed along the highways. Near real time information regarding certain roads' conditions are displayed by Waze and Google map, but, these data can only be used for monitoring traffic and travel navigation, alone, they cannot be used for highway modelling, planning, short-term traffic predictions and management. Only through calibrated highway macroscopic model, short term predictions can be obtained. This means when certain physical parameters of the highway are changed, the model can predict the traffic behaviour. Physical parameters are things like: number of lanes in main road, number of vehicles in one km/mile (density) at a certain time and location, number of lanes in off and on ramps. Short term prediction requires vehicle traffic data, such as traffic density, speed, or flow [4], and in turn it will provide useful data like: fundamental diagrams [5], speed-time-distance contour graphs, travel time [6] and distance travelled [7]. A proposed solution is by using nonintrusive traffic sensors only, namely smartphones-GPS enabled, with Virtual Detection Zone (VDZ) application and CCTV[8]. It has been proven through a few experiments that VDZ method combined with CCTV snap shots as a system, can potentially replace the traditional systems using inductive loop detector sensor. In order to achieve this goal, it is essential to show that automatic car counting can be obtained from the existing CCTV or better cameras, and by considering their F1 and precision [9] scores, give some recommendations. The counting methods investigated, firstly are the conventional ones like Back Subtraction (BS) and Viola Jones Algorithm (VJA) without or with Median (M) and Gaussian (G) Filters. Secondly, Deep Learning has also been attempted to count vehicles, either via static frames created from a video, or through progressing video frames.